Master Thesis
Attending courses and working at the Arbabian Lab on generative models for analog circuit design.
Hello ๐ I'm a Statistics & Data Science student at Ludwig Maximilian University of Munich (SIST scholar), focused on deep learning and language models. Previously founded Kaza Street Tech (street food order management in India), recognized in Shark Tank Campus Special Round and as a top student-led startup by Startup India.
Notes on deep learning, language models, and ML in production.
Loading latest posts...
Attending courses and working at the Arbabian Lab on generative models for analog circuit design.
Honors program run with LMU, TUM, and partners: about 25 students per cohort, team projects at the intersection of technology, product, and business. Alumni have founded 9+ unicorns; affiliated companies have raised more than โฌ9B.
Training infrastructure for computer-use agents: datasets of real computer-use trajectories and application-specific sandboxes for labs and AI teams.
Built an order management system for street food vendors.
Goal: ModernBERT extension for Indic languages. Prove the recipe on Hindi first, then scale to ModernBERT architectures for more than 5 Indic languages that each have substantial (>5B token) corpora available.
TL;DR: hindi-modernBERT is a pretrained base MLM checkpoint: a Hindi extension of the ModernBERT architecture, trained from scratch on ~29B Hindi text tokens. The base model is competitive with other models across tasks, and it outperforms them on retrieval after DPR fine-tuning. Checkpoint ba1157, 22 layers, 8192 context, ~188M params, trained on 1ร RTX 4090 in 5 days.
AI-assisted annotation and transcription for Byzantine Greek legal manuscripts, with Dr. Zachary Chitwood.
TL;DR: Nomicous is an AI-assisted annotation platform designed to make transcription of historical manuscripts 10ร faster. Models auto-segment and transcribe each page so humans only need to correct and review.
Building a PyTorch-like deep learning library from scratch (autograd, tensors, training loops) to deeply understand the fundamentals and use it to train Indic models for agentic use cases.
Distilling Geneformer (single-cell expression to network dynamics) into compact students at roughly 4.3M parameters, using far less data and compute while preserving ~83% of teacher classification performance.